Matīss commited on
Commit
a88cc29
1 Parent(s): 13d6772

Upload liv4ever.py

Browse files
Files changed (1) hide show
  1. liv4ever.py +139 -0
liv4ever.py ADDED
@@ -0,0 +1,139 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+ """Liv4ever dataset."""
16
+
17
+
18
+ import json
19
+
20
+ import datasets
21
+
22
+
23
+ _CITATION = """\
24
+ @inproceedings{rikters-etal-2022,
25
+ title = "Machine Translation for Livonian: Catering for 20 Speakers",
26
+ author = "Rikters, Matīss and
27
+ Tomingas, Marili and
28
+ Tuisk, Tuuli and
29
+ Valts, Ernštreits and
30
+ Fishel, Mark",
31
+ booktitle = "Proceedings of ACL 2022",
32
+ year = "2022",
33
+ address = "Dublin, Ireland",
34
+ publisher = "Association for Computational Linguistics"
35
+ }
36
+ """
37
+
38
+
39
+ _DESCRIPTION = """\
40
+ Livonian is one of the most endangered languages in Europe with just a tiny handful of speakers and virtually no publicly available corpora.
41
+ In this paper we tackle the task of developing neural machine translation (NMT) between Livonian and English, with a two-fold aim: on one hand,
42
+ preserving the language and on the other – enabling access to Livonian folklore, lifestories and other textual intangible heritage as well as
43
+ making it easier to create further parallel corpora. We rely on Livonian's linguistic similarity to Estonian and Latvian and collect parallel
44
+ and monolingual data for the four languages for translation experiments. We combine different low-resource NMT techniques like zero-shot translation,
45
+ cross-lingual transfer and synthetic data creation to reach the highest possible translation quality as well as to find which base languages are
46
+ empirically more helpful for transfer to Livonian. The resulting NMT systems and the collected monolingual and parallel data, including a manually
47
+ translated and verified translation benchmark, are publicly released.
48
+
49
+ Fields:
50
+ - source: source of the data
51
+ - en: sentence in English
52
+ - liv: sentence in Livonian
53
+ """
54
+
55
+ _HOMEPAGE = "https://huggingface.co/datasets/tartuNLP/liv4ever-data"
56
+
57
+ _LICENSE = "CC BY-NC-SA 4.0"
58
+
59
+ _REPO = "https://huggingface.co/datasets/tartuNLP/liv4ever/raw/main/"
60
+
61
+ _URLs = {
62
+ "train": _REPO + "train.json",
63
+ "dev": _REPO + "dev.json",
64
+ "test": _REPO + "test.json",
65
+ }
66
+
67
+
68
+ class liv4ever(datasets.GeneratorBasedBuilder):
69
+ """Liv4ever dataset."""
70
+
71
+ VERSION = datasets.Version("1.0.0")
72
+
73
+ def _info(self):
74
+ features = datasets.Features(
75
+ {
76
+ "source": datasets.Value("string"),
77
+ "en": datasets.Value("string"),
78
+ "liv: datasets.Value("string"),
79
+ }
80
+ )
81
+ return datasets.DatasetInfo(
82
+ description=_DESCRIPTION,
83
+ features=features,
84
+ supervised_keys=None,
85
+ homepage=_HOMEPAGE,
86
+ license=_LICENSE,
87
+ citation=_CITATION,
88
+ )
89
+
90
+ def _split_generators(self, dl_manager):
91
+ """Returns SplitGenerators."""
92
+ data_dir = dl_manager.download_and_extract(_URLs)
93
+
94
+ return [
95
+ datasets.SplitGenerator(
96
+ name=datasets.Split.TRAIN,
97
+ gen_kwargs={
98
+ "filepath": data_dir["train"],
99
+ "split": "train",
100
+ },
101
+ ),
102
+ datasets.SplitGenerator(
103
+ name=datasets.Split.TEST,
104
+ gen_kwargs={"filepath": data_dir["test"], "split": "test"},
105
+ ),
106
+ datasets.SplitGenerator(
107
+ name=datasets.Split.VALIDATION,
108
+ gen_kwargs={
109
+ "filepath": data_dir["dev"],
110
+ "split": "dev",
111
+ },
112
+ ),
113
+ ]
114
+
115
+ def _generate_examples(self, filepath, split):
116
+ """Yields examples."""
117
+
118
+ with open(filepath, encoding="utf-8") as f:
119
+ data = json.load(f)
120
+
121
+
122
+ for dialogue in data:
123
+ source = dialogue["source"]
124
+ sentences = dialogue["sentences"]
125
+
126
+ i=0
127
+
128
+ for turn in sentences:
129
+ i = i+1
130
+ sent_no = i
131
+ en = dialogue["en"]
132
+ liv = dialogue["liv"]
133
+
134
+ yield f"{sent_no}", {
135
+ "no": sent_no,
136
+ "source": source,
137
+ "en": en,
138
+ "liv": liv,
139
+ }